You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
rocksdb/table/iterator.cc

131 lines
4.2 KiB

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.
#include "rocksdb/iterator.h"
Eliminate unnecessary (slow) block cache Ref()ing in MultiGet (#9899) Summary: When MultiGet() determines that multiple query keys can be served by examining the same data block in block cache (one Lookup()), each PinnableSlice referring to data in that data block needs to hold on to the block in cache so that they can be released at arbitrary times by the API user. Historically this is accomplished with extra calls to Ref() on the Handle from Lookup(), with each PinnableSlice cleanup calling Release() on the Handle, but this creates extra contention on the block cache for the extra Ref()s and Release()es, especially because they hit the same cache shard repeatedly. In the case of merge operands (possibly more cases?), the problem was compounded by doing an extra Ref()+eventual Release() for each merge operand for a key reusing a block (which could be the same key!), rather than one Ref() per key. (Note: the non-shared case with `biter` was already one per key.) This change optimizes MultiGet not to rely on these extra, contentious Ref()+Release() calls by instead, in the shared block case, wrapping the cache Release() cleanup in a refcounted object referenced by the PinnableSlices, such that after the last wrapped reference is released, the cache entry is Release()ed. Relaxed atomic refcounts should be much faster than mutex-guarded Ref() and Release(), and much less prone to a performance cliff when MultiGet() does a lot of block sharing. Note that I did not use std::shared_ptr, because that would require an extra indirection object (shared_ptr itself new/delete) in order to associate a ref increment/decrement with a Cleanable cleanup entry. (If I assumed it was the size of two pointers, I could do some hackery to make it work without the extra indirection, but that's too fragile.) Some details: * Fixed (removed) extra block cache tracing entries in cases of cache entry reuse in MultiGet, but it's likely that in some other cases traces are missing (XXX comment inserted) * Moved existing implementations for cleanable.h from iterator.cc to new cleanable.cc * Improved API comments on Cleanable * Added a public SharedCleanablePtr class to cleanable.h in case others could benefit from the same pattern (potentially many Cleanables and/or smart pointers referencing a shared Cleanable) * Add a typedef for MultiGetContext::Mask * Some variable renaming for clarity Pull Request resolved: https://github.com/facebook/rocksdb/pull/9899 Test Plan: Added unit tests for SharedCleanablePtr. Greatly enhanced ability of existing tests to detect cache use-after-free. * Release PinnableSlices from MultiGet as they are read rather than in bulk (in db_test_util wrapper). * In ASAN build, default to using a trivially small LRUCache for block_cache so that entries are immediately erased when unreferenced. (Updated two tests that depend on caching.) New ASAN testsuite running time seems OK to me. If I introduce a bug into my implementation where we skip the shared cleanups on block reuse, ASAN detects the bug in `db_basic_test *MultiGet*`. If I remove either of the above testing enhancements, the bug is not detected. Consider for follow-up work: manipulate or randomize ordering of PinnableSlice use and release from MultiGet db_test_util wrapper. But in typical cases, natural ordering gives pretty good functional coverage. Performance test: In the extreme (but possible) case of MultiGetting the same or adjacent keys in a batch, throughput can improve by an order of magnitude. `./db_bench -benchmarks=multireadrandom -db=/dev/shm/testdb -readonly -num=5 -duration=10 -threads=20 -multiread_batched -batch_size=200` Before ops/sec, num=5: 1,384,394 Before ops/sec, num=500: 6,423,720 After ops/sec, num=500: 10,658,794 After ops/sec, num=5: 16,027,257 Also note that previously, with high parallelism, having query keys concentrated in a single block was worse than spreading them out a bit. Now concentrated in a single block is faster than spread out, which is hopefully consistent with natural expectation. Random query performance: with num=1000000, over 999 x 10s runs running before & after simultaneously (each -threads=12): Before: multireadrandom [AVG 999 runs] : 1088699 (± 7344) ops/sec; 120.4 (± 0.8 ) MB/sec After: multireadrandom [AVG 999 runs] : 1090402 (± 7230) ops/sec; 120.6 (± 0.8 ) MB/sec Possibly better, possibly in the noise. Reviewed By: anand1976 Differential Revision: D35907003 Pulled By: pdillinger fbshipit-source-id: bbd244d703649a8ca12d476f2d03853ed9d1a17e
2 years ago
#include "memory/arena.h"
#include "table/internal_iterator.h"
#include "table/iterator_wrapper.h"
namespace ROCKSDB_NAMESPACE {
Status Iterator::GetProperty(std::string prop_name, std::string* prop) {
if (prop == nullptr) {
return Status::InvalidArgument("prop is nullptr");
}
if (prop_name == "rocksdb.iterator.is-key-pinned") {
*prop = "0";
return Status::OK();
}
return Status::InvalidArgument("Unidentified property.");
}
namespace {
class EmptyIterator : public Iterator {
public:
explicit EmptyIterator(const Status& s) : status_(s) { }
bool Valid() const override { return false; }
void Seek(const Slice& /*target*/) override {}
void SeekForPrev(const Slice& /*target*/) override {}
void SeekToFirst() override {}
void SeekToLast() override {}
void Next() override { assert(false); }
void Prev() override { assert(false); }
Slice key() const override {
assert(false);
return Slice();
}
Slice value() const override {
assert(false);
return Slice();
}
Status status() const override { return status_; }
private:
Status status_;
};
template <class TValue = Slice>
class EmptyInternalIterator : public InternalIteratorBase<TValue> {
public:
explicit EmptyInternalIterator(const Status& s) : status_(s) {}
bool Valid() const override { return false; }
void Seek(const Slice& /*target*/) override {}
void SeekForPrev(const Slice& /*target*/) override {}
void SeekToFirst() override {}
void SeekToLast() override {}
void Next() override { assert(false); }
void Prev() override { assert(false); }
Slice key() const override {
assert(false);
return Slice();
}
TValue value() const override {
assert(false);
return TValue();
}
Status status() const override { return status_; }
private:
Status status_;
};
} // namespace
Iterator* NewEmptyIterator() { return new EmptyIterator(Status::OK()); }
Iterator* NewErrorIterator(const Status& status) {
return new EmptyIterator(status);
}
template <class TValue>
InternalIteratorBase<TValue>* NewErrorInternalIterator(const Status& status) {
return new EmptyInternalIterator<TValue>(status);
}
Add an option to put first key of each sst block in the index (#5289) Summary: The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes. Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it. So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks. Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files. This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289 Differential Revision: D15256423 Pulled By: al13n321 fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
template InternalIteratorBase<IndexValue>* NewErrorInternalIterator(
const Status& status);
template InternalIteratorBase<Slice>* NewErrorInternalIterator(
const Status& status);
template <class TValue>
InternalIteratorBase<TValue>* NewErrorInternalIterator(const Status& status,
Arena* arena) {
if (arena == nullptr) {
return NewErrorInternalIterator<TValue>(status);
} else {
auto mem = arena->AllocateAligned(sizeof(EmptyInternalIterator<TValue>));
return new (mem) EmptyInternalIterator<TValue>(status);
}
}
Add an option to put first key of each sst block in the index (#5289) Summary: The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes. Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it. So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks. Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files. This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289 Differential Revision: D15256423 Pulled By: al13n321 fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
template InternalIteratorBase<IndexValue>* NewErrorInternalIterator(
const Status& status, Arena* arena);
template InternalIteratorBase<Slice>* NewErrorInternalIterator(
const Status& status, Arena* arena);
template <class TValue>
InternalIteratorBase<TValue>* NewEmptyInternalIterator() {
return new EmptyInternalIterator<TValue>(Status::OK());
}
Add an option to put first key of each sst block in the index (#5289) Summary: The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes. Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it. So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks. Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files. This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289 Differential Revision: D15256423 Pulled By: al13n321 fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
template InternalIteratorBase<IndexValue>* NewEmptyInternalIterator();
template InternalIteratorBase<Slice>* NewEmptyInternalIterator();
template <class TValue>
InternalIteratorBase<TValue>* NewEmptyInternalIterator(Arena* arena) {
if (arena == nullptr) {
return NewEmptyInternalIterator<TValue>();
} else {
auto mem = arena->AllocateAligned(sizeof(EmptyInternalIterator<TValue>));
return new (mem) EmptyInternalIterator<TValue>(Status::OK());
}
}
Add an option to put first key of each sst block in the index (#5289) Summary: The first key is used to defer reading the data block until this file gets to the top of merging iterator's heap. For short range scans, most files never make it to the top of the heap, so this change can reduce read amplification by a lot sometimes. Consider the following workload. There are a few data streams (we'll be calling them "logs"), each stream consisting of a sequence of blobs (we'll be calling them "records"). Each record is identified by log ID and a sequence number within the log. RocksDB key is concatenation of log ID and sequence number (big endian). Reads are mostly relatively short range scans, each within a single log. Writes are mostly sequential for each log, but writes to different logs are randomly interleaved. Compactions are disabled; instead, when we accumulate a few tens of sst files, we create a new column family and start writing to it. So, a typical sst file consists of a few ranges of blocks, each range corresponding to one log ID (we use FlushBlockPolicy to cut blocks at log boundaries). A typical read would go like this. First, iterator Seek() reads one block from each sst file. Then a series of Next()s move through one sst file (since writes to each log are mostly sequential) until the subiterator reaches the end of this log in this sst file; then Next() switches to the next sst file and reads sequentially from that, and so on. Often a range scan will only return records from a small number of blocks in small number of sst files; in this case, the cost of initial Seek() reading one block from each file may be bigger than the cost of reading the actually useful blocks. Neither iterate_upper_bound nor bloom filters can prevent reading one block from each file in Seek(). But this PR can: if the index contains first key from each block, we don't have to read the block until this block actually makes it to the top of merging iterator's heap, so for short range scans we won't read any blocks from most of the sst files. This PR does the deferred block loading inside value() call. This is not ideal: there's no good way to report an IO error from inside value(). As discussed with siying offline, it would probably be better to change InternalIterator's interface to explicitly fetch deferred value and get status. I'll do it in a separate PR. Pull Request resolved: https://github.com/facebook/rocksdb/pull/5289 Differential Revision: D15256423 Pulled By: al13n321 fbshipit-source-id: 750e4c39ce88e8d41662f701cf6275d9388ba46a
5 years ago
template InternalIteratorBase<IndexValue>* NewEmptyInternalIterator(
Arena* arena);
template InternalIteratorBase<Slice>* NewEmptyInternalIterator(Arena* arena);
} // namespace ROCKSDB_NAMESPACE